tags
about

Was asked by a journalist to comment on the following article "DeepMind says reinforcement learning is ‘enough’ to reach general AI" and science paper "Reward is enough"

#Comment: This is classic Richard Sutton (one of the co-authors, a pioneer or RL), who has been reinforcing such views for 30+ years. From my perspective, this is clearly a somewhat fringe philosophical position, misleadingly presented as "hard science" (computer science/math, etc).

In the general context of this paper, the scientific community is aware of countless very challenging problems and "known unknowns" (and even more "unknown unknowns"), which rightfully instill a sense of humility in most practitioners in this field and prevent them to make such grandiose, totalitarian statements ("RL is the final answer, all you need is reward"). Yet in somewhat typical deepmind fashion, they chose to make a bold statements that grabs (media) attention at all cost, over a more nuanced approach. This is more akin to politics than science.

But make no mistake, Paul Dirac's statement "There are two main problems in A.I: What is A and what is I" is still very much a productive characterization of where we are at in AGI research. The only thing that has fundamentally changed since the 1950/60s, is that Science-Fiction is now a valid tool for giant corporations to willingly confuse and mislead the public, journalists and shareholders.

#ML